KameleonLabs

Blog / Article

The model ‘too capable to release’

Anthropic introduces Claude Mythos Preview, its most powerful model, and decides not to make it publicly available: it reserves it for a group of companies and organizations to find and fix security flaws. The decision opens a new debate about who gets access to the most advanced AI.

#Security and risks #Models and releases

Apr 13, 2026 · 2 min read

A technology company introduces its best product and, in the same breath, announces that it will not sell it to the public. It sounds like a paradox, but it is exactly what has just happened. And the reason is not commercial, but security-related.

A launch without a launch

On Tuesday, April 7, Anthropic introduced Claude Mythos Preview, its most capable model to date, alongside Project Glasswing, an initiative to protect critical software. Instead of opening it to everyone, it has placed it in the hands of twelve partners, including Amazon Web Services, Apple, Google, Microsoft, Nvidia, CrowdStrike, JPMorganChase and the Linux Foundation, and more than 40 organizations that maintain software infrastructure. The company is contributing up to $100 million in usage credits and says it has no plans for a general release.

Flaws hidden for decades

The reason is how well the model finds vulnerabilities: programming errors that an attacker could exploit. According to Anthropic, it has already detected thousands of serious flaws in the major operating systems and browsers, including a 27-year-old one in OpenBSD, a system known precisely for its security. The same ability that helps plug holes could also help exploit them, and the company fears it could fall into hands that would not use it carefully.

The same key that checks every lock could also open them.

The alarm reaches banking

The issue has quickly moved beyond the technical world. According to Bloomberg and Reuters, Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell brought together the chief executives of several major US banks this week to make sure they understand the risks posed by models of this kind. AI cybersecurity being treated as a potential risk to the financial system is a development in itself.

Caution or an exclusive club?

Reactions are divided. Some see an act of responsibility: giving defenders a head start so they can fix flaws before similar models become available to anyone. Others raise uncomfortable questions: should a private company decide who gets access to technology like this? What about small and medium-sized businesses, hospitals or public administrations that are not on the list? Some also point out that, without independent evaluation, it is hard to judge from the outside how much real risk there is.

Opening up carefully

Our reading is that this episode marks a new stage: the most advanced models are no longer simply released, but also administered. That requires clear rules on access, transparency and external oversight, which barely exist today. But it also opens a real opportunity: if AI can quickly find flaws that have been hidden for decades, the software we all use could become considerably safer. For any organization, the practical message is to review its digital hygiene now, because the security bar is rising for everyone.

Related

ServiceSoftware developmentServiceConsultingTechnology foundationSquadwise